Core ConceptsOversight

Oversight

Four primitives decide and prove; this one keeps a person in the loop where a person belongs. Oversight is not a dashboard. It is the set of moments where Regent refuses to act on an agent’s word alone and asks the human who answers for it, and the record it keeps of what that human said.

Where a person is required

MomentWhat happensWhat the record says
Binding an agentevery agent is registered under a responsible party: a verified person or an organization’s verified administratorresponsible_party_id on the agent, anchored with its identity
Granting authoritylimits are terms. A mandate is activated only by the owner’s signature over them: a passkey in a production organization, the session in a sandbox; an API key cannot activate a mandateapproval_method, approval_digest in every token and receipt under that mandate
Escalationa payment above the ask-first threshold, a medium behavioural risk on a money action, a policy rule that says ask: the agent stops, the owner sees the request (console, e-mail, a card in the chat), approves or declines with the passkey; the agent retries the same call with the same idempotency_keythe decision carries the approval; the receipt carries approval_method
Step-upabove a threshold the organization sets, a passkey is not enough: the approver confirms with their face at the identity providerapproval_method: passkey+authid; no biometric is stored by Regent
Asking for morean agent may ask for a budget, up to a ceiling; only a person can grant itthe request lands under Approvals and in the owners’ inbox; nothing changes until signed
Stoppingthe kill switch: revoking an agent suspends every mandate it holds within milliseconds; deleting the owner’s keys at the identity provider or an expired identity document stops new authority the same wayAGENT_NOT_ACTIVE, MANDATE_SUSPENDED on the next call; the revocation in the audit trail

What the console is

The console (Agents, Mandates, Approvals, Activity, Settings) is where a person does these things. Everything it shows is also in the API and the audit trail; everything it signs is signed with the person’s own passkey, so the console never holds authority of its own. Agents reach the same moments without the console: an escalation card in the chat, an e-mail link, regent-control onboard for an agent with a terminal.

Why it is a primitive

A control plane that only enforces limits is a calculator. The limits are only worth enforcing if a person set them knowingly, and an action outside them is only safe if a person decided it. Oversight is what turns the system allowed it into the owner allowed it, which is the sentence an auditor, an insurer or a counterparty actually needs.